Deep Learning Based Object Shape Identification from EOG Controlled Vision System

摘要

Humans look at the object of interest prior to grasping it. This phenomenon leads to necessity of artificial vision system integrated with prosthetic hand for object shape identification. Individuals who use prosthetic arms are able to move their eyes to locate objects. The eye movements can easily be interpreted by Electrooculography (EOG) signal. In this study, a single webcam placed in a cap visor, synchronised to move in the same direction as per the user's eye gaze angle. The camera clicked the snapshot including the object immediately fixing the eye over the object. To recognise the object shape, Convolutional Neural Network (CNN) was used further. Using VGG16 architecture, objects belonging to whole hand grasps were classified into cylindrical, spherical, cubical and conical types to adhere to most of the daily used objects. Around 93.0% accuracy was achieved with the realtime objects or object views that were not included in the training set.